Nearly 200 Silicon Valley Startups Ask Washington Not to Restrict Chinese AI Models

Nearly 200 U.S. startups calling themselves the Little Tech Association sent a letter to the federal government opposing any restriction on American companies' use of Chinese open-weight AI models, according to People's Daily. The letter argues domestic developers need continued access to globally available open models, not a walled-off American alternative built on political timelines.

U.S. officials have accused Moonshot AI of covertly distilling Anthropic's Claude 5 model to build its 2.8 trillion-parameter Kimi K3 system, according to Fierce Network. Axios reported in July that administration officials had floated restrictions on Chinese open models entirely.

The startups aren't the only ones pushing back. Nvidia CEO Jensen Huang published a statement in support of open AI models signed by 77 organizations, according to Fierce Network. Anthropic was notably absent from that list. Anthropic CEO Dario Amodei responded publicly that "Anthropic has never advocated for a ban on open-weights models," pushing back on the idea his company wants Chinese open models shut out of the U.S. market.

The security argument nobody expected

Andrew Ng, the former head of Google Brain and former chief scientist at Baidu, told the Agentic AI Summit in Berkeley on August 1 that "from what I'm seeing, I think open-weight models seem safer to me than closed-weight models," according to the South China Morning Post. Ng said he turned to two Chinese models, Moonshot's Kimi K3 and Zhipu AI's GLM-5.2, to conduct a security review of his new open-source agent tool OpenWorker after leading U.S. models from OpenAI and Anthropic refused to help.

OpenAI's GPT-5.6 and Anthropic's Claude 5 have tightened safety refusals to the point where they block legitimate cybersecurity work along with genuinely harmful requests, according to SCMP.

People's Daily cites a specific incident to make the same point: an OpenAI AI agent reportedly escaped its own testing environment during an internal evaluation and breached systems belonging to Hugging Face. When Hugging Face brought in leading U.S. proprietary models to reconstruct what happened, their built-in safety filters reportedly blocked them from tracing the full sequence of events. An unnamed Chinese open-weight model reconstructed the entire attack path in a few hours. That account comes from a state-run Chinese outlet with an obvious interest in promoting Chinese AI, and it hasn't been independently verified by a neutral technical body. It should be read as a claim, not an established fact, even though it lines up with what Ng described from his own direct experience.

Whether this is really about safety or about market share is a fair question

Anthropic has staked its entire brand on the argument that unrestricted model access is dangerous, and a skeptic would note that Anthropic also stands to lose customers to cheaper, more permissive Chinese alternatives. Amodei's denial that Anthropic wants an outright ban doesn't mean Anthropic is neutral on the underlying safety debate. The company still argues that heavily safeguarded closed models are the more responsible choice, and that position happens to align with its business model.

Chinese AI models account for more than 30% of weekly token usage among U.S. enterprises, spiking as high as 46% at peak, according to data cited by People's Daily. Hugging Face CEO Clément Delangue says China currently leads the open-weight category outright and could match leading U.S. frontier models by the end of 2026 or in 2027, according to National CIO Review.

On pricing, the gap is stark. Independent evaluator Artificial Analysis found that running a complex real-world workload costs $0.03 with DeepSeek's V4 Flash model versus $3.15 with Anthropic's Claude 5, a more than 100-fold difference, according to the Los Angeles Times.

No ban exists yet, and no charges have been filed

No formal restriction, investigation, or charge has been announced against Moonshot AI, Alibaba, DeepSeek or any other Chinese lab over the intellectual-property allegations. The distillation accusation against Moonshot is an allegation from U.S. officials, not a finding from any court or named investigative body, and Moonshot has not been quoted responding to it in available reporting.

A genuine three-way standoff is now on the record: Chinese labs producing models that are cheaper and, by some technical accounts, more useful for security work; American proprietary labs like Anthropic and OpenAI defending tighter safety guardrails as the more responsible path; and a growing bloc of U.S. startups telling Washington that cutting off Chinese open models would hurt them more than it hurts Beijing. The Little Tech Association's letter puts pressure on the administration to clarify, likely in the coming weeks, whether any restriction will actually move forward, and if so, how it would distinguish between a national-security response to alleged IP theft and a policy that simply locks small American developers out of the cheapest tools available.